首页|Learnable confdence‑driven asymmetric attention fusion mechanism for PET/CT tumor segmentation

Learnable confdence‑driven asymmetric attention fusion mechanism for PET/CT tumor segmentation

扫码查看
原文链接
  • NETL
  • NSTL
  • Springer Nature
Positron emission tomography/computed tomography (PET/CT) imaging, combin- ing PET’s sensitivity with CT’s resolution, is pivotal in clinical tumor screening. PET/CT tumor segmentation forms the basis for treatment planning and surgical guidance. Recent advancements in deep learning-based PET/CT tumor segmenta- tion, notably via feature fusion strategies, have shown promise. However, existing fusion strategies have not accounted for the phenomenon where certain tumors are more prominent in PET images while others are more prominent in CT images, thus limiting their ability to fully exploit features from both modalities. To address this, we propose an learnable confdence-driven asymmetric attention fusion (LCAAF) mechanism. Initially, we train PET and CT segmentation branches separately and propose learnable segmentation confdence to evaluate single-modality segmenta- tion without labels. Leveraging this confdence, we design an asymmetric attention mechanism that integrates encoder features from both modalities into the decoder. The proposed mechanism enhances the contribution of the modality with higher confdence in feature fusion, thereby improving segmentation outcomes. Experi- mental results on soft tissue sarcoma (STS) and headNeck datasets demonstrate the efcacy of assigning higher weights to modalities with prominent tumor regions during feature fusion. The proposed method yields a 1.87% higher Dice value com- pared to the highest-performing CSAE-Net model on the STS dataset and a 4.88% higher Dice value compared to the highest CSAE-Net model on the HeadNeck data- set. Additionally, the proposed segmentation confdence can serve as an evaluation metric for label-free segmentation results.

Biomedical segmentationPET/CTConfdenceAttention

Zhaoshuo Diao、Yan Zhang、Yang Xu、Xi Chen、Zongzhe Sun

展开 >

School of Software,Shenyang University of Technology,Shenyang 110870,China

Institute for Prevention and Control of AIDS and Sexually Transmitted Diseases,Liaoning Provincial Center for Disease Prevention and Control,Shenyang 110005,China||Software College,Northeastern University,Shenyang 110819,China

People's Hospital of Mudan,Heze 274000,China

China Academy of Information and Communications Technology,Beijing 100083,China

展开 >

2025

10.1007/s11227-025-07313-8

The Journal of Supercomputing

The Journal of Supercomputing

SCI
ISSN：0920-8542
年,卷(期)：2025.81(7)
  • 34